
!pip install imblearn

Requirement already satisfied: imblearn in c:\users\aritr\anaconda3\lib\site-packages (0.0)
Requirement already satisfied: imbalanced-learn in c:\users\aritr\anaconda3\lib\site-packages (from imblearn) (0.10.1)
Requirement already satisfied: joblib>=1.1.1 in c:\users\aritr\anaconda3\lib\site-packages (from imbalanced-learn->imblearn) (1.2.0)
Requirement already satisfied: threadpoolctl>=2.0.0 in c:\users\aritr\anaconda3\lib\site-packages (from imbalanced-learn->imblearn) (2.2.0)
Requirement already satisfied: scikit-learn>=1.0.2 in c:\users\aritr\anaconda3\lib\site-packages (from imbalanced-learn->imblearn) (1.0.2)
Requirement already satisfied: numpy>=1.17.3 in c:\users\aritr\anaconda3\lib\site-packages (from imbalanced-learn->imblearn) (1.21.5)
Requirement already satisfied: scipy>=1.3.2 in c:\users\aritr\anaconda3\lib\site-packages (from imbalanced-learn->imblearn) (1.7.3)


import pandas as pd
import numpy as np
import warnings
warnings.filterwarnings("ignore")
import seaborn as sns
import matplotlib.pyplot as plt 
from datetime import datetime
from sklearn.model_selection import train_test_split
from statsmodels.stats.outliers_influence import variance_inflation_factor as vif
from sklearn.model_selection import GridSearchCV
from sklearn import metrics
from sklearn.linear_model import LogisticRegression
from sklearn.neighbors import KNeighborsClassifier
from sklearn.metrics import accuracy_score
from sklearn.metrics import confusion_matrix
from sklearn.metrics import classification_report
from sklearn.metrics import roc_auc_score, roc_curve
from imblearn.over_sampling import SMOTE
from sklearn.model_selection import cross_validate
from sklearn.model_selection import RepeatedStratifiedKFold
import plotly.express as px


df=pd.read_csv('ml_data.csv')
df


df.isna().sum()

Invoice             0
StockCode           0
Description      4194
Quantity            0
InvoiceDate         0
Price               0
Customer ID    211901
Country             0
dtype: int64


# Since prediction will be for customer ids, removing records having no customer ID values
df2=df[df['Customer ID'].notna()].reset_index().drop('index',axis=1)
df2


df2.isna().sum()

Invoice        0
StockCode      0
Description    0
Quantity       0
InvoiceDate    0
Price          0
Customer ID    0
Country        0
dtype: int64


#Check for if any Customer ID has decimal values other than 0

# df3=df2.copy()
# df3['dummy']=df3['Customer ID']
# df3['dummy']=df3['dummy'].astype(int)
# df3
# df3[df3['Customer ID']>df3.dummy]


#Changing data type of Customer ID to string

df2['Customer ID']=df2['Customer ID'].astype(int)
df2['Customer ID']=df2['Customer ID'].astype(str)

df2['InvoiceDate']=pd.to_datetime(df2['InvoiceDate'])
df2.dtypes

Invoice                object
StockCode              object
Description            object
Quantity                int64
InvoiceDate    datetime64[ns]
Price                 float64
Customer ID            object
Country                object
dtype: object


# Statistical summary of numerical data - understanding data ranges 
df2.describe()


# Understanding reason behind negative values of price - if cancelled orders only have negative quantity
x=df2[df2.Invoice.str.contains('C')]
print(len(x[x.Quantity<0]),len(x))
y=df2[df2.Quantity>=0]
y[y.Invoice.str.contains('C')]

17341 17341


# Getting net Customer Spend : Calculate Price * Quantity for every Record
df2['Net_Cust_Spend']=df2['Price']*df2['Quantity']
df2['Net_Cust_Spend_mod']=np.where(df2['Net_Cust_Spend'] <0,0,df2['Net_Cust_Spend'])
df2


# dataframe split based on time
## ADS (Modelling) : 2009-12-01 to 2011-10-09
## Test Set : 2011-10-10 to 2011-11-09

#filtering for records beyond 1 month of max InvoiceDate value present in dataset
th_dt=datetime.strptime('2011-11-09',"%Y-%m-%d")+pd.Timedelta(days=-30)
ads_df=df2[df2.InvoiceDate<th_dt]
ads_df

#Get max InvoiceDate for every customer in unfiltered data
cust_invdt_ttl=df2.groupby(['Customer ID'],as_index=False).agg({'InvoiceDate':'max'}).rename({
    'InvoiceDate':'InvoiceDate_max'},axis=1)
cust_invdt_ttl

#Get common customers in modelling ADS and unfiltered data
ads=pd.DataFrame({'Customer ID':list(ads_df['Customer ID'].unique())})
ads=pd.merge(ads,cust_invdt_ttl,how='left',on='Customer ID')
ads

#Check for customers buying products within 2011-10-10 and 2011-11-09, flag them as 1, rest as 0 -> target column
ads['next_mnth_prchse']=np.where((ads['InvoiceDate_max']>=th_dt),1,0)
ads=ads[['Customer ID','next_mnth_prchse']]
ads


# Distribution of target column - class imbalance found
ads['next_mnth_prchse'].value_counts()

0    4263
1    1309
Name: next_mnth_prchse, dtype: int64


dist_df=ads.groupby(['next_mnth_prchse'],as_index=False).size().rename({'size':'count_class'},axis=1)
dist_df['pct_count']=np.round(dist_df['count_class']*100/len(ads),2)
dist_df['pct_count']=dist_df['pct_count'].apply(lambda x:str(x)+' %')
dist_df
bar=px.bar(dist_df,x='next_mnth_prchse',y='count_class',text='pct_count',title='Target Value Distribution')
bar.show()


# Recency - R
## logic : For every customer Id - Calculate difference in days between 2011-10-09 and their last purchase date
ads_cust_invdt=ads_df.groupby(['Customer ID'],as_index=False).agg({'InvoiceDate':'max'}).rename({
    'InvoiceDate':'InvoiceDate_max'},axis=1)
ads_cust_invdt['Max_dt']=th_dt+pd.Timedelta(days=-1)
ads_cust_invdt['Recency']=(ads_cust_invdt['Max_dt']-ads_cust_invdt['InvoiceDate_max']).dt.days

# Since only dates are subtracted and time isnt considered here, matching max_dt and InvoiceDate_max values 
# will have diff < 0. So they are replaced by 0 Due to match in Date
ads_cust_invdt['Recency']=np.where(ads_cust_invdt['Recency']<0,0,ads_cust_invdt['Recency'])
ads_cust_invdt=ads_cust_invdt[['Customer ID','Recency']]

# Add Recency to main ads
ads=pd.merge(ads_cust_invdt,ads,how='inner',on='Customer ID')
ads


# Frequency - F
## Logic : For every Customer ID, get count of distrinct InvoiceDate values
ads_freq=ads_df.groupby(['Customer ID'],as_index=False).agg({'InvoiceDate':'nunique'}).rename({
    'InvoiceDate':'Frequency'},axis=1)

# Adding Frequency data to main ads
ads=pd.merge(ads_freq,ads,how='inner',on='Customer ID')
ads


# Monetary Value / Revenue - M
## Logic : For every Customer ID, get sum of entire Net_Cust_Spend
ads_rev=ads_df.groupby(['Customer ID'],as_index=False).agg({'Net_Cust_Spend_mod':'sum'}).rename({
    'Net_Cust_Spend_mod':'Totl_Money_Spent'},axis=1)

# Add Monetary value to main ads
ads=pd.merge(ads_rev,ads,how='inner',on='Customer ID')
ads


# Checking distribution of R,F,M values to understand data spread and range
ads[['Totl_Money_Spent','Frequency','Recency']].describe()


#Transformation of RFM values
ads['Totl_Money_Spent_trf']=(ads['Totl_Money_Spent']-ads['Totl_Money_Spent'].min())/(ads['Totl_Money_Spent'].max()
                                                                                     -ads['Totl_Money_Spent'].min())
ads['Frequency_trf']=(ads['Frequency']-ads['Frequency'].min())/(ads['Frequency'].max()-ads['Frequency'].min())
ads['Recency_trf']=(ads['Recency']-ads['Recency'].min())/(ads['Recency'].max()-ads['Recency'].min())
ads


ads[['Totl_Money_Spent_trf','Frequency_trf','Recency_trf']].describe()


# Getting Recency scores
ads['Recency_Score']=pd.qcut(ads['Recency_trf'],q=4,labels=[4,3,2,1])
ads


# Getting Frequency scores
ads['Frequency_Score']=pd.qcut(ads['Frequency_trf'].rank(method='first'),q=4,labels=[1,2,3,4])
ads


# Getting Revenue scores
ads['Totl_Money_Spent_Score']=pd.qcut(ads['Totl_Money_Spent_trf'],q=4,labels=[1,2,3,4])
ads


# Multiply weights with RFM scores
ads['w_Recency_Score']=ads['Recency_Score'].astype(int)*0.2
ads['w_Frequency_Score']=ads['Frequency_Score'].astype(int)*0.5
ads['w_Totl_Money_Spent_Score']=ads['Totl_Money_Spent_Score'].astype(int)*0.3
ads


# Calculating Net score for all RFM scores combined
ads['Net_RFM_Score']=ads['w_Recency_Score']+ads['w_Frequency_Score']+ads['w_Totl_Money_Spent_Score']
ads


ads_final=ads[['Customer ID','Net_RFM_Score','next_mnth_prchse']]
ads_final


# Split into train and valid set
Y=ads_final[['next_mnth_prchse']]
X=ads_final.drop('next_mnth_prchse',axis=1)
x_train,x_valid,y_train,y_valid=train_test_split(X,Y,test_size=0.3, stratify=Y)
x_train


# Removing Customer ID from train and valid data
x_train_mod=x_train.drop('Customer ID',axis=1)
x_valid_mod=x_valid.drop('Customer ID',axis=1)


# # Checking for multicollinearity
# vif_df = pd.DataFrame()
# vif_df["feature"] = x_train_vif.columns
  
# # calculating VIF for each feature
# vif_df["VIF"] = [vif(x_train_vif.values, i)
#                           for i in range(len(x_train_vif.columns))]

# vif_df


# Initializing logistic regression
classifier_lr = LogisticRegression(random_state=300,multi_class='ovr')
classifier_lr.fit(x_train_mod, y_train)

LogisticRegression(multi_class='ovr', random_state=300)


#Predictions for logistic regression and buy-probability calculation - valid set
y_pred_valid_lr = classifier_lr.predict(x_valid_mod)
y_pred_valid_lr=pd.Series(y_pred_valid_lr)
pred_valid_prob_lr=classifier_lr.predict_proba(x_valid_mod)[:, 1]


#Predictions for logistic regression and buy-probability calculation - train set
y_pred_train_lr = classifier_lr.predict(x_train_mod)
y_pred_train_lr=pd.Series(y_pred_train_lr)


def model_performance(y_test,y_pred,opt):
    # Accuracy Score
    print ("Accuracy ",opt," : ", accuracy_score(y_test, y_pred),'\n')
        
    # Confusion Matrix
    cm = confusion_matrix(y_test, y_pred)
    ax= plt.subplot()
    sns.heatmap(cm, annot=True, fmt='g', ax=ax)  #annot=True to annotate cells, ftm='g' to disable scientific notation
    ax.set_xlabel('Predicted labels')
    ax.set_ylabel('True labels')
    ax.set_title('Confusion Matrix')
    ax.xaxis.set_ticklabels(['won"t buy(0)','will buy(1)'])
    ax.yaxis.set_ticklabels(['won"t buy(0)','will buy(1)'])
    
    # Classification report
    target_names = ['wont buy(0)', 'will buy(1)']
    print(classification_report(y_test, y_pred, target_names=target_names))


model_performance(y_valid,y_pred_valid_lr,'valid set')

Accuracy  valid set  :  0.7828947368421053 

              precision    recall  f1-score   support

 wont buy(0)       0.81      0.94      0.87      1279
 will buy(1)       0.58      0.27      0.37       393

    accuracy                           0.78      1672
   macro avg       0.70      0.60      0.62      1672
weighted avg       0.75      0.78      0.75      1672


model_performance(y_train,y_pred_train_lr,'train set')

Accuracy  train set  :  0.7923076923076923 

              precision    recall  f1-score   support

 wont buy(0)       0.81      0.95      0.87      2984
 will buy(1)       0.62      0.29      0.40       916

    accuracy                           0.79      3900
   macro avg       0.72      0.62      0.64      3900
weighted avg       0.77      0.79      0.76      3900


# ROC-AUC
auc_lr = roc_auc_score(y_valid, pred_valid_prob_lr)
fpr_lr, tpr_lr, thresholds_lr = roc_curve(y_valid, pred_valid_prob_lr)

plt.figure(figsize=(5, 5))
plt.plot(fpr_lr, tpr_lr, label=f'AUC (Logistic Regression) = {auc_lr:.2f}')
plt.plot([0, 1], [0, 1], color='blue', linestyle='--', label='Baseline')
plt.title('ROC Curve', size=20)
plt.xlabel('False Positive Rate', size=14)
plt.ylabel('True Positive Rate', size=14)
plt.legend()

<matplotlib.legend.Legend at 0x23295310be0>


rec_c1 = []
target_names = ['wont buy(0)', 'will buy(1)']
for i in range(1,10):
    neigh = KNeighborsClassifier(n_neighbors = i).fit(x_train_mod, y_train)
    yhat = neigh.predict(x_valid_mod)
    cr=classification_report(y_valid, yhat, target_names=target_names,output_dict=True)
    rec_c1.append(cr['will buy(1)']['recall'])
    
print("Maximum recall of class 1 :-",max(rec_c1),"at K =",rec_c1.index(max(rec_c1))+1)

Maximum recall of class 1 :- 0.37404580152671757 at K = 3


k=rec_c1.index(max(rec_c1))+1
classifier_knn = KNeighborsClassifier(n_neighbors = k).fit(x_train_mod, y_train)


#Predictions for logistic regression and buy-probability calculation - valid set
y_pred_valid_knn = classifier_knn.predict(x_valid_mod)
y_pred_valid_knn=pd.Series(y_pred_valid_knn)
pred_valid_prob_knn=classifier_knn.predict_proba(x_valid_mod)[:, 1]


#Predictions for logistic regression and buy-probability calculation - train set
y_pred_train_knn = classifier_knn.predict(x_train_mod)
y_pred_train_knn=pd.Series(y_pred_train_knn)


model_performance(y_valid,y_pred_valid_knn,'valid set')

Accuracy  valid set  :  0.743421052631579 

              precision    recall  f1-score   support

 wont buy(0)       0.82      0.86      0.84      1279
 will buy(1)       0.45      0.37      0.41       393

    accuracy                           0.74      1672
   macro avg       0.63      0.62      0.62      1672
weighted avg       0.73      0.74      0.74      1672


model_performance(y_train,y_pred_train_knn,'train set')

Accuracy  train set  :  0.7576923076923077 

              precision    recall  f1-score   support

 wont buy(0)       0.82      0.87      0.85      2984
 will buy(1)       0.48      0.39      0.43       916

    accuracy                           0.76      3900
   macro avg       0.65      0.63      0.64      3900
weighted avg       0.74      0.76      0.75      3900


# ROC-AUC
auc_knn = roc_auc_score(y_valid, pred_valid_prob_knn)
fpr_knn, tpr_knn, thresholds_knn = roc_curve(y_valid, pred_valid_prob_knn)

plt.figure(figsize=(5, 5))
plt.plot(fpr_knn, tpr_knn, label=f'AUC (KNN Classifier) = {auc_knn:.2f}')
plt.plot([0, 1], [0, 1], color='blue', linestyle='--', label='Baseline')
plt.title('ROC Curve', size=20)
plt.xlabel('False Positive Rate', size=14)
plt.ylabel('True Positive Rate', size=14)
plt.legend()

<matplotlib.legend.Legend at 0x232954da790>


# Making a copy of final ADS
X_2=X.copy()
Y_2=Y.copy()
X_2=X_2.drop('Customer ID',axis=1)

# Increase occurances of class 1
ov_sp=SMOTE()
ov_X,ov_Y=ov_sp.fit_resample(X_2, Y_2)

#train-valid data split
ov_x_train, ov_x_valid, ov_y_train, ov_y_valid = train_test_split(ov_X, ov_Y, test_size=0.3, stratify=ov_Y)
ov_x_train


# Initializing logistic regression
smote_classifier_lr = LogisticRegression(random_state=300,multi_class='ovr')
smote_classifier_lr.fit(ov_x_train,ov_y_train)

LogisticRegression(multi_class='ovr', random_state=300)


# apply smote rf model to predict for non-smote valid set
y_pred_valid_lrsmote=smote_classifier_lr.predict(x_valid_mod)
y_pred_valid_lrsmote=pd.Series(y_pred_valid_lrsmote)
pred_valid_prob_lrsmote=smote_classifier_lr.predict_proba(x_valid_mod)[:, 1]


# apply smote rf model to predict for non-smote train set
y_pred_train_lrsmote=smote_classifier_lr.predict(x_train_mod)
y_pred_train_lrsmote=pd.Series(y_pred_train_lrsmote)


model_performance(y_valid,y_pred_valid_lrsmote,'valid set')

Accuracy  valid set  :  0.6561004784688995 

              precision    recall  f1-score   support

 wont buy(0)       0.87      0.65      0.74      1279
 will buy(1)       0.37      0.68      0.48       393

    accuracy                           0.66      1672
   macro avg       0.62      0.66      0.61      1672
weighted avg       0.75      0.66      0.68      1672


model_performance(y_train,y_pred_train_lrsmote,'train set')

Accuracy  train set  :  0.6684615384615384 

              precision    recall  f1-score   support

 wont buy(0)       0.88      0.66      0.75      2984
 will buy(1)       0.39      0.70      0.50       916

    accuracy                           0.67      3900
   macro avg       0.63      0.68      0.63      3900
weighted avg       0.76      0.67      0.69      3900


# ROC-AUC for Smote RF results
auc_lr = roc_auc_score(y_valid, pred_valid_prob_lrsmote)
fpr_lr, tpr_lr, thresholds_lr = roc_curve(y_valid, pred_valid_prob_lrsmote)

plt.figure(figsize=(7, 7))
plt.plot(fpr_lr, tpr_lr, label=f'AUC (SMOTE Logistic Regression Classification) = {auc_lr:.2f}')
plt.plot([0, 1], [0, 1], color='blue', linestyle='--', label='Baseline')
plt.title('ROC Curve', size=20)
plt.xlabel('False Positive Rate', size=14)
plt.ylabel('True Positive Rate', size=14)
plt.legend()

<matplotlib.legend.Legend at 0x23295675760>


#Create Stratified K-fold cross validation to validate model performance and check for variance
cv = RepeatedStratifiedKFold(n_splits=10, n_repeats=3, random_state=300)
scoring = ('f1', 'recall', 'precision','accuracy')
#Evaluate SMOTE RF model
scores = cross_validate(smote_classifier_lr, ov_X, ov_Y, scoring=scoring, cv=cv)
#Get average evaluation metrics

print('Mean f1: %.3f' % np.mean(scores['test_f1']))
print('Mean recall: %.3f' % np.mean(scores['test_recall']))
print('Mean precision: %.3f' % np.mean(scores['test_precision']))
print('Mean accuracy: %.3f' % np.mean(scores['test_accuracy']))

Mean f1: 0.682
Mean recall: 0.696
Mean precision: 0.669
Mean accuracy: 0.676


# Test Set Preparation : Create Dataframe of all Customer IDs till 2011-11-09
test_set=df2.groupby(['Customer ID'],as_index=False).size()
test_set=test_set[['Customer ID']]
test_set


# Calculating RFM values for test set
# Recency - R
ads_cust_invdt=df2.groupby(['Customer ID'],as_index=False).agg({'InvoiceDate':'max'}).rename({
    'InvoiceDate':'InvoiceDate_max'},axis=1)
ads_cust_invdt['Max_dt']=df2['InvoiceDate'].max()
ads_cust_invdt['Recency']=(ads_cust_invdt['Max_dt']-ads_cust_invdt['InvoiceDate_max']).dt.days

# In case of negative date difference, it will be replaced by 0
ads_cust_invdt['Recency']=np.where(ads_cust_invdt['Recency']<0,0,ads_cust_invdt['Recency'])
ads_cust_invdt=ads_cust_invdt[['Customer ID','Recency']]
ads_cust_invdt

# Add Recency to main ads
test_set=pd.merge(ads_cust_invdt,test_set,how='inner',on='Customer ID')


# Frequency - F
ads_freq=df2.groupby(['Customer ID'],as_index=False).agg({'InvoiceDate':'nunique'}).rename({
    'InvoiceDate':'Frequency'},axis=1)

# Adding Frequency data to main ads
test_set=pd.merge(ads_freq,test_set,how='inner',on='Customer ID')


# Monetary Value / Revenue - M
## Logic : For every Customer ID, get sum of entire Net_Cust_Spend
ads_rev=df2.groupby(['Customer ID'],as_index=False).agg({'Net_Cust_Spend_mod':'sum'}).rename({
    'Net_Cust_Spend_mod':'Totl_Money_Spent'},axis=1)

# Add Monetary value to main ads
test_set=pd.merge(ads_rev,test_set,how='inner',on='Customer ID')
test_set


# Applying data transformation on RFM values
test_set['Totl_Money_Spent_trf']=(test_set['Totl_Money_Spent']-test_set['Totl_Money_Spent'].min())/(
    test_set['Totl_Money_Spent'].max()-test_set['Totl_Money_Spent'].min())
test_set['Frequency_trf']=(test_set['Frequency']-test_set['Frequency'].min())/(test_set['Frequency'].max()-
                                                                               test_set['Frequency'].min())
test_set['Recency_trf']=(test_set['Recency']-test_set['Recency'].min())/(test_set['Recency'].max()-
                                                                         test_set['Recency'].min())
test_set


test_set.describe()


test_set['Recency_Score']=pd.qcut(test_set['Recency_trf'],q=4,labels=[4,3,2,1])
test_set['Frequency_Score']=pd.qcut(test_set['Frequency_trf'].rank(method='first'),q=4,labels=[1,2,3,4])
test_set['Totl_Money_Spent_Score']=pd.qcut(test_set['Totl_Money_Spent_trf'],q=4,labels=[1,2,3,4])
test_set


test_set['w_Recency_Score']=test_set['Recency_Score'].astype(int)*0.2
test_set['w_Frequency_Score']=test_set['Frequency_Score'].astype(int)*0.5
test_set['w_Totl_Money_Spent_Score']=test_set['Totl_Money_Spent_Score'].astype(int)*0.3

test_set['Net_RFM_Score']=test_set['w_Recency_Score']+test_set['w_Frequency_Score']+test_set['w_Totl_Money_Spent_Score']
test_set


# Selecting final set of features in test set
test_set_final=test_set[['Customer ID']+list(ov_x_train.columns)]
test_set_final


# Excluding Customer ID for prediction
test_pred_df=test_set_final.drop('Customer ID',axis=1)
test_pred_df


# Predicting probability of customers buying any product in the require dtime period (next month)
pred_final=smote_classifier_lr.predict_proba(test_pred_df)[:, 1]
pred_final=pd.DataFrame(pred_final).rename({0:'Prediction'},axis=1)
pred_final


# Adding Prediction data back to main test set
sub=pd.merge(test_set_final,pred_final,left_index=True,right_index=True,how='inner')
sub=sub[['Customer ID','Prediction']]
sub


#  Getting % of customers who can be potential buyers next month
len(sub[sub.Prediction>0.5])*100/len(sub)

42.6643598615917

	Invoice	StockCode	Description	Quantity	InvoiceDate	Price	Customer ID	Country
0	489434	85048	15CM CHRISTMAS GLASS BALL 20 LIGHTS	12	2009-12-01 07:45:00	6.95	13085.0	United Kingdom
1	489434	79323P	PINK CHERRY LIGHTS	12	2009-12-01 07:45:00	6.75	13085.0	United Kingdom
2	489434	79323W	WHITE CHERRY LIGHTS	12	2009-12-01 07:45:00	6.75	13085.0	United Kingdom
3	489434	22041	RECORD FRAME 7" SINGLE SIZE	48	2009-12-01 07:45:00	2.10	13085.0	United Kingdom
4	489434	21232	STRAWBERRY CERAMIC TRINKET BOX	24	2009-12-01 07:45:00	1.25	13085.0	United Kingdom
...	...	...	...	...	...	...	...	...
944463	575312	22083	PAPER CHAIN KIT RETROSPOT	6	2011-11-09 12:49:00	2.95	13588.0	United Kingdom
944464	575312	23355	HOT WATER BOTTLE KEEP CALM	4	2011-11-09 12:49:00	4.95	13588.0	United Kingdom
944465	575312	22110	BIRD HOUSE HOT WATER BOTTLE	6	2011-11-09 12:49:00	2.55	13588.0	United Kingdom
944466	575312	22037	ROBOT BIRTHDAY CARD	12	2011-11-09 12:49:00	0.42	13588.0	United Kingdom
944467	575312	21790	VINTAGE SNAP CARDS	12	2011-11-09 12:49:00	0.85	13588.0	United Kingdom

	Invoice	StockCode	Description	Quantity	InvoiceDate	Price	Customer ID	Country
0	489434	85048	15CM CHRISTMAS GLASS BALL 20 LIGHTS	12	2009-12-01 07:45:00	6.95	13085.0	United Kingdom
1	489434	79323P	PINK CHERRY LIGHTS	12	2009-12-01 07:45:00	6.75	13085.0	United Kingdom
2	489434	79323W	WHITE CHERRY LIGHTS	12	2009-12-01 07:45:00	6.75	13085.0	United Kingdom
3	489434	22041	RECORD FRAME 7" SINGLE SIZE	48	2009-12-01 07:45:00	2.10	13085.0	United Kingdom
4	489434	21232	STRAWBERRY CERAMIC TRINKET BOX	24	2009-12-01 07:45:00	1.25	13085.0	United Kingdom
...	...	...	...	...	...	...	...	...
732562	575312	22083	PAPER CHAIN KIT RETROSPOT	6	2011-11-09 12:49:00	2.95	13588.0	United Kingdom
732563	575312	23355	HOT WATER BOTTLE KEEP CALM	4	2011-11-09 12:49:00	4.95	13588.0	United Kingdom
732564	575312	22110	BIRD HOUSE HOT WATER BOTTLE	6	2011-11-09 12:49:00	2.55	13588.0	United Kingdom
732565	575312	22037	ROBOT BIRTHDAY CARD	12	2011-11-09 12:49:00	0.42	13588.0	United Kingdom
732566	575312	21790	VINTAGE SNAP CARDS	12	2011-11-09 12:49:00	0.85	13588.0	United Kingdom

	Quantity	Price
count	732567.000000	732567.000000
mean	12.792626	3.766560
std	147.626721	74.355886
min	-74215.000000	0.000000
25%	2.000000	1.250000
50%	5.000000	1.950000
75%	12.000000	3.750000
max	74215.000000	38970.000000

	Invoice	StockCode	Description	Quantity	InvoiceDate	Price	Customer ID	Country	Net_Cust_Spend	Net_Cust_Spend_mod
0	489434	85048	15CM CHRISTMAS GLASS BALL 20 LIGHTS	12	2009-12-01 07:45:00	6.95	13085	United Kingdom	83.40	83.40
1	489434	79323P	PINK CHERRY LIGHTS	12	2009-12-01 07:45:00	6.75	13085	United Kingdom	81.00	81.00
2	489434	79323W	WHITE CHERRY LIGHTS	12	2009-12-01 07:45:00	6.75	13085	United Kingdom	81.00	81.00
3	489434	22041	RECORD FRAME 7" SINGLE SIZE	48	2009-12-01 07:45:00	2.10	13085	United Kingdom	100.80	100.80
4	489434	21232	STRAWBERRY CERAMIC TRINKET BOX	24	2009-12-01 07:45:00	1.25	13085	United Kingdom	30.00	30.00
...	...	...	...	...	...	...	...	...	...	...
732562	575312	22083	PAPER CHAIN KIT RETROSPOT	6	2011-11-09 12:49:00	2.95	13588	United Kingdom	17.70	17.70
732563	575312	23355	HOT WATER BOTTLE KEEP CALM	4	2011-11-09 12:49:00	4.95	13588	United Kingdom	19.80	19.80
732564	575312	22110	BIRD HOUSE HOT WATER BOTTLE	6	2011-11-09 12:49:00	2.55	13588	United Kingdom	15.30	15.30
732565	575312	22037	ROBOT BIRTHDAY CARD	12	2011-11-09 12:49:00	0.42	13588	United Kingdom	5.04	5.04
732566	575312	21790	VINTAGE SNAP CARDS	12	2011-11-09 12:49:00	0.85	13588	United Kingdom	10.20	10.20

	Customer ID	next_mnth_prchse
0	13085	0
1	13078	1
2	15362	0
3	18102	1
4	12682	1
...	...	...
5567	16685	0
5568	13193	0
5569	13212	0
5570	14801	0
5571	16380	0

Null Treatment¶

Data Preprocessing¶

ADS preparation¶

Calculating RFM values to understand proirity of customers for the store¶

RFM Scores¶

Weighted RFM scoring¶

Modelling Pipeline¶

Logistic Regression¶

KNN Classifier¶

Restructured Modelling Pipeline with SMOTE application¶

SMOTE Logistic Regression Classifier¶

Probability Prediction for customers to buy a product between 2011-11-09 to 2011-12-09¶

Calculating RFM values¶

Getting RFM scores¶

Applying weights on scores¶

Prediction on test set¶

	Customer ID	Recency	next_mnth_prchse
0	12346	263	0
1	12347	67	1
2	12348	13	0
3	12349	345	0
4	12350	248	0
...	...	...	...
5567	18283	33	1
5568	18284	367	0
5569	18285	598	0
5570	18286	414	0
5571	18287	139	1

	Customer ID	Totl_Money_Spent	Frequency	Recency	next_mnth_prchse
0	12346	77556.46	17	263	0
1	12347	3402.39	6	67	1
2	12348	2019.40	5	13	0
3	12349	2671.14	4	345	0
4	12350	334.40	1	248	0
...	...	...	...	...	...
5567	18283	1706.20	16	33	1
5568	18284	461.68	2	367	0
5569	18285	427.00	1	598	0
5570	18286	1296.43	3	414	0
5571	18287	3110.99	6	139	1

	Totl_Money_Spent	Frequency	Recency
count	5572.000000	5572.000000	5572.000000
mean	2687.265253	7.003410	205.079684
std	13033.889368	14.503232	189.960525
min	0.000000	1.000000	0.000000
25%	313.635000	1.000000	30.000000
50%	776.730000	3.000000	144.000000
75%	2099.940000	7.250000	344.000000
max	524476.600000	431.000000	676.000000

	Customer ID	Totl_Money_Spent	Frequency	Recency	next_mnth_prchse	Totl_Money_Spent_trf	Frequency_trf	Recency_trf
0	12346	77556.46	17	263	0	0.147874	0.037209	0.389053
1	12347	3402.39	6	67	1	0.006487	0.011628	0.099112
2	12348	2019.40	5	13	0	0.003850	0.009302	0.019231
3	12349	2671.14	4	345	0	0.005093	0.006977	0.510355
4	12350	334.40	1	248	0	0.000638	0.000000	0.366864
...	...	...	...	...	...	...	...	...
5567	18283	1706.20	16	33	1	0.003253	0.034884	0.048817
5568	18284	461.68	2	367	0	0.000880	0.002326	0.542899
5569	18285	427.00	1	598	0	0.000814	0.000000	0.884615
5570	18286	1296.43	3	414	0	0.002472	0.004651	0.612426
5571	18287	3110.99	6	139	1	0.005932	0.011628	0.205621

	Customer ID	Net_RFM_Score
4667	17311	3.5
3642	16231	2.8
5114	17787	3.0
1652	14109	1.6
4682	17327	1.5
...	...	...
4428	17058	3.2
3906	16511	3.0
2182	14671	2.1
2735	15258	2.2
4900	17559	1.0

	Net_RFM_Score
4338	1.5
7183	3.3
5363	1.7
2596	1.0
7109	3.2
...	...
395	3.8
4059	2.4
3295	3.2
1308	2.5
3067	4.0

	Customer ID	Totl_Money_Spent	Frequency	Recency	Totl_Money_Spent_trf	Frequency_trf	Recency_trf
0	12346	77556.46	17	295	0.136252	0.034409	0.416667
1	12347	4696.71	7	9	0.008251	0.012903	0.012712
2	12348	2019.40	5	44	0.003548	0.008602	0.062147
3	12349	2671.14	4	377	0.004693	0.006452	0.532486
4	12350	334.40	1	279	0.000587	0.000000	0.394068
...	...	...	...	...	...	...	...
5775	18283	1819.19	17	12	0.003196	0.034409	0.016949
5776	18284	461.68	2	399	0.000811	0.002151	0.563559
5777	18285	427.00	1	630	0.000750	0.000000	0.889831
5778	18286	1296.43	3	446	0.002278	0.004301	0.629944
5779	18287	4182.99	8	12	0.007349	0.015054	0.016949

	Totl_Money_Spent	Frequency	Recency	Totl_Money_Spent_trf	Frequency_trf	Recency_trf
count	5780.000000	5780.000000	5780.000000	5780.000000	5780.000000	5780.000000
mean	2773.135119	7.176817	203.424048	0.004872	0.013283	0.287322
std	13702.082997	15.023021	201.343859	0.024072	0.032308	0.284384
min	0.000000	1.000000	0.000000	0.000000	0.000000	0.000000
25%	320.535000	1.000000	28.000000	0.000563	0.000000	0.039548
50%	810.985000	3.000000	124.500000	0.001425	0.004301	0.175847
75%	2124.025000	8.000000	362.000000	0.003731	0.015054	0.511299
max	569215.420000	466.000000	708.000000	1.000000	1.000000	1.000000

	Prediction
0	0.694500
1	0.673811
2	0.561344
3	0.537640
4	0.233268
...	...
5775	0.714433
5776	0.288509
5777	0.288509
5778	0.465936
5779	0.769294